CSRConstraint: use label columns; frozen constraints break after later variable add/remove

[FabianHofmann]

Note

The following content was generated by AI (Claude Code), reviewed by @FabianHofmann.

Bug: frozen constraints break after any later variable change

CSRConstraint stores dense variable positions as column indices and a fixed column count (n_active_vars at freeze time). Any add_variables or remove_variables after freezing changes both, so model.matrices fails when stacking the blocks:

ValueError: inconsistent shapes (3, 6) and (1, 8)

This is independent of #870; it reproduces on master with a plain dense freeze=True constraint.

Reproduction

import linopy, pandas as pd

m = linopy.Model()
i = pd.RangeIndex(3, name="i")
a = m.add_variables(coords=[i], name="a")
b = m.add_variables(coords=[i], name="b")
m.add_constraints(2 * b >= 1, name="c1", freeze=True)

d = m.add_variables(coords=[pd.RangeIndex(2, name="i")], name="d")
m.add_constraints(d <= 5, name="c2", freeze=True)
m.matrices.A  # ValueError: inconsistent shapes (3, 6) and (1, 8)

# same after m.remove_variables("a") instead of adding: (3, 6) vs (1, 5)

Proposal: label columns for CSRConstraint

Store raw variable labels as column indices (width model._xCounter), and map labels to dense positions once at matrix assembly via VariableLabelIndex.label_to_pos. This is the convention CSRPayload (#870) already uses for expressions, so the two CSR representations would agree.

Consequences:

1. to_matrix / to_matrix_with_rhs gather positions through label_to_pos and return n_active_vars-wide blocks; the bug above disappears by construction. has_labels, to_polars and the dataset reconstruction no longer need vlabels lookups.
2. from_payload drops its label-to-position mapping and no longer needs the label index at construction time.
3. The netcdf _linopy_format: csr writer stores labels; the reader gets a compatibility path for files written with positions (they carry vlabels in the same file).
4. LinearExpression.to_constraint on a payload-backed expression can return an unassigned CSRConstraint (cindex=None) directly. Constraint._from_pending, the lazy Constraint.data branch and extract_pending introduced in feat(v1): CSR-backed sparse groupby-sum - skew-independent build memory (6-9x on the #745 hub case) #870 as a stopgap can then be deleted; add_constraints only assigns labels.
5. linopy/persistent/snapshot.py reads to_matrix_with_rhs output and needs a check that buffer ownership assumptions still hold.

This is also the prerequisite for making the CSR-backed constraint the default without a dense-to-sparse conversion step.

Refs: #870